DEEP NEURAL NETWORK MODEL FOR SEARCHING ARTICLE IN THE BOOK OF INDONESIAN CRIMINAL LAW

• Main Author: Fiqih Caesandria, Novinda
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/46749
• Institution: Institut Teknologi Bandung

The number of legal regulations makes it difficult for people to understand the rules. Many regulations can be understood easily by using a search technique for law books in Indonesian. The technique of searching for regulations, especially in the Criminal Code previously had been carried out. This research tries to make a search on the criminal law book with a method that is currently developing, namely deep neural network. The search for KUHP article uses a query in the form of cases of violations of law recorded at the Supreme Court. This research applies a search with sentence pair modeling using the deep neural network method BiLSTM and CNN. Vector representation for each word feature is trained using word embedding and there is an entity extraction process with NER. In testing the deep neural network method compared to the Vector Space Model method, which is LSI that was previously implemented in article search and Elasticsearch. The results of this research test show that the deep neural network approach has better results than the frequency vector technique that has been tried using LSI with a mean reciprocal rank value of 0.056 in the top 5 ranks. The performance in the top 5 rank of BiLSTM model with the mean reciprocal rank is 0.2558, while for CNN is 0.1379. The mean reciprocal rank of elasticsearch is 0.3032 in the top 5, so the results of the proposed method do not produce better performance than elasticsearch. That is because, several words with a same meaning in word embedding have a small similarity value. Some cases in this study also cannot overcome compound sentences. The addition of NER in this study can add performance.